50. CNN - yojulab/learn_deeplearning GitHub Wiki

Convolutional Neural Networks (CNNs / ConvNets)

  • ๊ธฐ์กด Fully Connect ๋ฐฉ์‹ ํ•œ๊ณ„ : ์ด๋ฏธ์ง€ ๋ถ„๋ฅ˜์—” ์•ฝํ•จ.
  • ๋ถ„๋ฅ˜ ์—๋Ÿฌ์œจ

MLP vs CNN

  • NN
  • CNN

์‚ฌ์ „ ์ดํ•ด

์ฃผ์š” ๊ตฌ์„ฑ Layers

Convolution Layer

?dilated convolution

Pooling Layer

  • ์ด๋ฏธ์ง€ ํ•ด์ƒ๋„ ๋‚ฎ์ถ”๋Š” ๋ชฉ์  : ์˜์ƒ ํ•„ํ„ฐ๋ง ์ด์šฉ
  • Feature ์ˆ˜๋ฅผ ์ถ•์†Œ(์š”์•ฝ)ํ•˜์—ฌ ์ถ”์ƒํ™” ํ•˜๋Š” ํšจ๊ณผ
  • ๋งŽ์€ feature ์ˆ˜๋Š” Overfitting ๊ฒฝํ–ฅ ๋ฐœ์ƒ
  • General pooling

Up-Sampling Layer

  • Pooling Layer ๋ฐ˜๋Œ€ ์—ญํ• 
  • ์ด๋ฏธ์ง€ ํ•ด์ƒ๋„ ํ–ฅ์ƒ ๋ชฉ์ 

์ฃผ์š” ํŠน์ง•

  • Shift Invariance : ์ด๋ฏธ์ง€ ๋‚ด ์‚ฌ๋ฌผ ์œ„์น˜ ๊ด€๊ณ„ ์—†์ด ์ธ์‹ ๊ฐ€๋Šฅ
  • Layer๊ฐ€ ๊นŠ์–ด์งˆ์ˆ˜๋ก high-level feature๋กœ ์ •๋ณด ์ƒ์„ฑ
  • ipynb-CIFAR10 simple, ipynb - CIFAR10

์ฃผ์š” Models

Playground